[PATCH] D87788: [DAGCombiner] Teach visitMLOAD to replace an all ones mask with an unmasked load

[Craig Topper via Phabricator via llvm-commits]

This revision was automatically updated to reflect the committed changes.
Closed by commit rG89ee4c0314bd: [DAGCombiner] Teach visitMLOAD to replace an all ones mask with an unmasked load (authored by craig.topper).

Changed prior to commit:
  https://reviews.llvm.org/D87788?vs=292311&id=292328#toc

Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D87788/new/

https://reviews.llvm.org/D87788

Files:
  llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
  llvm/test/CodeGen/X86/masked_load.ll


Index: llvm/test/CodeGen/X86/masked_load.ll
===================================================================
--- llvm/test/CodeGen/X86/masked_load.ll
+++ llvm/test/CodeGen/X86/masked_load.ll
@@ -6171,25 +6171,10 @@
 ; SSE-NEXT:    movups (%rdi), %xmm0
 ; SSE-NEXT:    retq
 ;
-; AVX1OR2-LABEL: mload_constmask_v4f32_all:
-; AVX1OR2:       ## %bb.0:
-; AVX1OR2-NEXT:    vmovups (%rdi), %xmm0
-; AVX1OR2-NEXT:    retq
-;
-; AVX512F-LABEL: mload_constmask_v4f32_all:
-; AVX512F:       ## %bb.0:
-; AVX512F-NEXT:    movw $15, %ax
-; AVX512F-NEXT:    kmovw %eax, %k1
-; AVX512F-NEXT:    vmovups (%rdi), %zmm0 {%k1} {z}
-; AVX512F-NEXT:    ## kill: def $xmm0 killed $xmm0 killed $zmm0
-; AVX512F-NEXT:    vzeroupper
-; AVX512F-NEXT:    retq
-;
-; AVX512VL-LABEL: mload_constmask_v4f32_all:
-; AVX512VL:       ## %bb.0:
-; AVX512VL-NEXT:    kxnorw %k0, %k0, %k1
-; AVX512VL-NEXT:    vmovups (%rdi), %xmm0 {%k1} {z}
-; AVX512VL-NEXT:    retq
+; AVX-LABEL: mload_constmask_v4f32_all:
+; AVX:       ## %bb.0:
+; AVX-NEXT:    vmovups (%rdi), %xmm0
+; AVX-NEXT:    retq
   %res = call <4 x float> @llvm.masked.load.v4f32.p0v4f32(<4 x float>* %addr, i32 4, <4 x i1> <i1 true, i1 true, i1 true, i1 true>, <4 x float>undef)
   ret <4 x float> %res
 }
@@ -6573,7 +6558,7 @@
   ret <8 x double> %res
 }
 
-; FIXME: We should be able to detect the mask is all ones after type
+; Make sure we detect the mask is all ones after type
 ; legalization to use an unmasked load for some of the avx512 instructions.
 define <16 x double> @mload_constmask_v16f64_allones_split(<16 x double>* %addr, <16 x double> %dst) {
 ; SSE-LABEL: mload_constmask_v16f64_allones_split:
@@ -6611,29 +6596,26 @@
 ;
 ; AVX512F-LABEL: mload_constmask_v16f64_allones_split:
 ; AVX512F:       ## %bb.0:
-; AVX512F-NEXT:    kxnorw %k0, %k0, %k1
-; AVX512F-NEXT:    vmovupd (%rdi), %zmm0 {%k1}
 ; AVX512F-NEXT:    movb $85, %al
 ; AVX512F-NEXT:    kmovw %eax, %k1
 ; AVX512F-NEXT:    vmovupd 64(%rdi), %zmm1 {%k1}
+; AVX512F-NEXT:    vmovups (%rdi), %zmm0
 ; AVX512F-NEXT:    retq
 ;
 ; AVX512VLDQ-LABEL: mload_constmask_v16f64_allones_split:
 ; AVX512VLDQ:       ## %bb.0:
-; AVX512VLDQ-NEXT:    kxnorw %k0, %k0, %k1
-; AVX512VLDQ-NEXT:    vmovupd (%rdi), %zmm0 {%k1}
 ; AVX512VLDQ-NEXT:    movb $85, %al
 ; AVX512VLDQ-NEXT:    kmovw %eax, %k1
 ; AVX512VLDQ-NEXT:    vmovupd 64(%rdi), %zmm1 {%k1}
+; AVX512VLDQ-NEXT:    vmovups (%rdi), %zmm0
 ; AVX512VLDQ-NEXT:    retq
 ;
 ; AVX512VLBW-LABEL: mload_constmask_v16f64_allones_split:
 ; AVX512VLBW:       ## %bb.0:
-; AVX512VLBW-NEXT:    kxnorw %k0, %k0, %k1
-; AVX512VLBW-NEXT:    vmovupd (%rdi), %zmm0 {%k1}
 ; AVX512VLBW-NEXT:    movb $85, %al
 ; AVX512VLBW-NEXT:    kmovd %eax, %k1
 ; AVX512VLBW-NEXT:    vmovupd 64(%rdi), %zmm1 {%k1}
+; AVX512VLBW-NEXT:    vmovups (%rdi), %zmm0
 ; AVX512VLBW-NEXT:    retq
   %res = call <16 x double> @llvm.masked.load.v16f64.p0v16f64(<16 x double>* %addr, i32 4, <16 x i1> <i1 1, i1 1, i1 1, i1 1, i1 1, i1 1, i1 1, i1 1, i1 1, i1 0, i1 1, i1 0, i1 1, i1 0, i1 1, i1 0>, <16 x double> %dst)
   ret <16 x double> %res
Index: llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
===================================================================
--- llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
+++ llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
@@ -9272,6 +9272,16 @@
   if (ISD::isBuildVectorAllZeros(Mask.getNode()))
     return CombineTo(N, MLD->getPassThru(), MLD->getChain());
 
+  // If this is a masked load with an all ones mask, we can use a unmasked load.
+  // FIXME: Can we do this for indexed, expanding, or extending loads?
+  if (ISD::isBuildVectorAllOnes(Mask.getNode()) &&
+      MLD->isUnindexed() && !MLD->isExpandingLoad() &&
+      MLD->getExtensionType() == ISD::NON_EXTLOAD) {
+    SDValue NewLd = DAG.getLoad(N->getValueType(0), SDLoc(N), MLD->getChain(),
+                                MLD->getBasePtr(), MLD->getMemOperand());
+    return CombineTo(N, NewLd, NewLd.getValue(1));
+  }
+
   // Try transforming N to an indexed load.
   if (CombineToPreIndexedLoadStore(N) || CombineToPostIndexedLoadStore(N))
     return SDValue(N, 0);


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D87788.292328.patch
Type: text/x-patch
Size: 4100 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20200916/b929be39/attachment.bin>